Nvidia CEO Jensen Huang stated at the GTC conference in Berlin that the operational lifespan of GPUs has extended from the traditional four years to eight years or longer. This shift is driven by continuous software updates and the expanding demand for AI applications.
Huang noted that Nvidia's programmable software approach allows for ongoing performance improvements throughout the hardware lifecycle. This ensures that computing systems remain efficient and durable for extended periods, providing a significant advantage for users across various industries.
Elon Musk supported this view during a SpaceX earnings call, suggesting that the value of compute per watt increases over time. As AI models become more efficient, they generate more useful intelligence per unit of power, which keeps the value of older hardware like the H100 high.
Gavin Baker, an expert in the semiconductor supply chain, observed that while the cost per token is decreasing, the rental cost per GPU hour has risen significantly. This trend reflects the increased utility and value extracted from each hour of GPU processing time as software optimization improves.
